static NLboolean __nlInvert_SUPERLU(__NLContext *context) {
/* OpenNL Context */
NLfloat* b = (context->least_squares)? context->Mtb: context->b;
NLfloat* x = context->x;
NLuint n = context->n, j;
/* SuperLU variables */
SuperMatrix B;
NLint info;
for(j=0; j<context->nb_rhs; j++, b+=n, x+=n) {
/* Create superlu array for B */
sCreate_Dense_Matrix(
&B, n, 1, b, n,
SLU_DN, /* Fortran-type column-wise storage */
SLU_S, /* floats */
SLU_GE /* general */
);
/* Forward/Back substitution to compute x */
sgstrs(TRANS, &(context->slu.L), &(context->slu.U),
context->slu.perm_c, context->slu.perm_r, &B,
&(context->slu.stat), &info);
if(info == 0)
memcpy(x, ((DNformat*)B.Store)->nzval, sizeof(*x)*n);
Destroy_SuperMatrix_Store(&B);
}
return (info == 0);
}
int main(int argc, char *argv[])
{
void smatvec_mult(float alpha, float x[], float beta, float y[]);
void spsolve(int n, float x[], float y[]);
extern int sfgmr( int n,
void (*matvec_mult)(float, float [], float, float []),
void (*psolve)(int n, float [], float[]),
float *rhs, float *sol, double tol, int restrt, int *itmax,
FILE *fits);
extern int sfill_diag(int n, NCformat *Astore);
char equed[1] = {'B'};
yes_no_t equil;
trans_t trans;
SuperMatrix A, L, U;
SuperMatrix B, X;
NCformat *Astore;
NCformat *Ustore;
SCformat *Lstore;
GlobalLU_t Glu; /* facilitate multiple factorizations with
SamePattern_SameRowPerm */
float *a;
int *asub, *xa;
int *etree;
int *perm_c; /* column permutation vector */
int *perm_r; /* row permutations from partial pivoting */
int nrhs, ldx, lwork, info, m, n, nnz;
float *rhsb, *rhsx, *xact;
float *work = NULL;
float *R, *C;
float u, rpg, rcond;
float zero = 0.0;
float one = 1.0;
mem_usage_t mem_usage;
superlu_options_t options;
SuperLUStat_t stat;
FILE *fp = stdin;
int restrt, iter, maxit, i;
double resid;
float *x, *b;
#ifdef DEBUG
extern int num_drop_L, num_drop_U;
#endif
#if ( DEBUGlevel>=1 )
CHECK_MALLOC("Enter main()");
#endif
/* Defaults */
lwork = 0;
nrhs = 1;
trans = NOTRANS;
/* Set the default input options:
options.Fact = DOFACT;
options.Equil = YES;
options.ColPerm = COLAMD;
options.DiagPivotThresh = 0.1; //different from complete LU
options.Trans = NOTRANS;
options.IterRefine = NOREFINE;
options.SymmetricMode = NO;
options.PivotGrowth = NO;
options.ConditionNumber = NO;
options.PrintStat = YES;
options.RowPerm = LargeDiag;
options.ILU_DropTol = 1e-4;
options.ILU_FillTol = 1e-2;
options.ILU_FillFactor = 10.0;
options.ILU_DropRule = DROP_BASIC | DROP_AREA;
options.ILU_Norm = INF_NORM;
options.ILU_MILU = SILU;
*/
ilu_set_default_options(&options);
/* Modify the defaults. */
options.PivotGrowth = YES; /* Compute reciprocal pivot growth */
options.ConditionNumber = YES;/* Compute reciprocal condition number */
if ( lwork > 0 ) {
work = SUPERLU_MALLOC(lwork);
if ( !work ) ABORT("Malloc fails for work[].");
}
/* Read matrix A from a file in Harwell-Boeing format.*/
if (argc < 2)
{
printf("Usage:\n%s [OPTION] < [INPUT] > [OUTPUT]\nOPTION:\n"
"-h -hb:\n\t[INPUT] is a Harwell-Boeing format matrix.\n"
"-r -rb:\n\t[INPUT] is a Rutherford-Boeing format matrix.\n"
"-t -triplet:\n\t[INPUT] is a triplet format matrix.\n",
argv[0]);
return 0;
}
else
{
switch (argv[1][1])
{
case 'H':
case 'h':
printf("Input a Harwell-Boeing format matrix:\n");
sreadhb(fp, &m, &n, &nnz, &a, &asub, &xa);
break;
case 'R':
case 'r':
printf("Input a Rutherford-Boeing format matrix:\n");
sreadrb(&m, &n, &nnz, &a, &asub, &xa);
break;
case 'T':
case 't':
printf("Input a triplet format matrix:\n");
sreadtriple(&m, &n, &nnz, &a, &asub, &xa);
break;
default:
printf("Unrecognized format.\n");
return 0;
}
}
sCreate_CompCol_Matrix(&A, m, n, nnz, a, asub, xa,
SLU_NC, SLU_S, SLU_GE);
Astore = A.Store;
sfill_diag(n, Astore);
printf("Dimension %dx%d; # nonzeros %d\n", A.nrow, A.ncol, Astore->nnz);
fflush(stdout);
/* Generate the right-hand side */
if ( !(rhsb = floatMalloc(m * nrhs)) ) ABORT("Malloc fails for rhsb[].");
if ( !(rhsx = floatMalloc(m * nrhs)) ) ABORT("Malloc fails for rhsx[].");
sCreate_Dense_Matrix(&B, m, nrhs, rhsb, m, SLU_DN, SLU_S, SLU_GE);
sCreate_Dense_Matrix(&X, m, nrhs, rhsx, m, SLU_DN, SLU_S, SLU_GE);
xact = floatMalloc(n * nrhs);
ldx = n;
sGenXtrue(n, nrhs, xact, ldx);
sFillRHS(trans, nrhs, xact, ldx, &A, &B);
if ( !(etree = intMalloc(n)) ) ABORT("Malloc fails for etree[].");
if ( !(perm_r = intMalloc(m)) ) ABORT("Malloc fails for perm_r[].");
if ( !(perm_c = intMalloc(n)) ) ABORT("Malloc fails for perm_c[].");
if ( !(R = (float *) SUPERLU_MALLOC(A.nrow * sizeof(float))) )
ABORT("SUPERLU_MALLOC fails for R[].");
if ( !(C = (float *) SUPERLU_MALLOC(A.ncol * sizeof(float))) )
ABORT("SUPERLU_MALLOC fails for C[].");
info = 0;
#ifdef DEBUG
num_drop_L = 0;
num_drop_U = 0;
#endif
/* Initialize the statistics variables. */
StatInit(&stat);
/* Compute the incomplete factorization and compute the condition number
and pivot growth using dgsisx. */
B.ncol = 0; /* not to perform triangular solution */
sgsisx(&options, &A, perm_c, perm_r, etree, equed, R, C, &L, &U, work,
lwork, &B, &X, &rpg, &rcond, &Glu, &mem_usage, &stat, &info);
/* Set RHS for GMRES. */
if (!(b = floatMalloc(m))) ABORT("Malloc fails for b[].");
if (*equed == 'R' || *equed == 'B') {
for (i = 0; i < n; ++i) b[i] = rhsb[i] * R[i];
} else {
for (i = 0; i < m; i++) b[i] = rhsb[i];
}
printf("sgsisx(): info %d, equed %c\n", info, equed[0]);
if (info > 0 || rcond < 1e-8 || rpg > 1e8)
printf("WARNING: This preconditioner might be unstable.\n");
if ( info == 0 || info == n+1 ) {
if ( options.PivotGrowth == YES )
printf("Recip. pivot growth = %e\n", rpg);
if ( options.ConditionNumber == YES )
printf("Recip. condition number = %e\n", rcond);
} else if ( info > 0 && lwork == -1 ) {
printf("** Estimated memory: %d bytes\n", info - n);
}
Lstore = (SCformat *) L.Store;
Ustore = (NCformat *) U.Store;
printf("n(A) = %d, nnz(A) = %d\n", n, Astore->nnz);
printf("No of nonzeros in factor L = %d\n", Lstore->nnz);
printf("No of nonzeros in factor U = %d\n", Ustore->nnz);
printf("No of nonzeros in L+U = %d\n", Lstore->nnz + Ustore->nnz - n);
printf("Fill ratio: nnz(F)/nnz(A) = %.3f\n",
((double)(Lstore->nnz) + (double)(Ustore->nnz) - (double)n)
/ (double)Astore->nnz);
printf("L\\U MB %.3f\ttotal MB needed %.3f\n",
mem_usage.for_lu/1e6, mem_usage.total_needed/1e6);
fflush(stdout);
/* Set the global variables. */
GLOBAL_A = &A;
GLOBAL_L = &L;
GLOBAL_U = &U;
GLOBAL_STAT = &stat;
GLOBAL_PERM_C = perm_c;
GLOBAL_PERM_R = perm_r;
GLOBAL_OPTIONS = &options;
GLOBAL_R = R;
GLOBAL_C = C;
GLOBAL_MEM_USAGE = &mem_usage;
/* Set the options to do solve-only. */
options.Fact = FACTORED;
options.PivotGrowth = NO;
options.ConditionNumber = NO;
/* Set the variables used by GMRES. */
restrt = SUPERLU_MIN(n / 3 + 1, 50);
maxit = 1000;
iter = maxit;
resid = 1e-8;
if (!(x = floatMalloc(n))) ABORT("Malloc fails for x[].");
if (info <= n + 1)
{
int i_1 = 1;
double maxferr = 0.0, nrmA, nrmB, res, t;
float temp;
extern float snrm2_(int *, float [], int *);
extern void saxpy_(int *, float *, float [], int *, float [], int *);
/* Initial guess */
for (i = 0; i < n; i++) x[i] = zero;
t = SuperLU_timer_();
/* Call GMRES */
sfgmr(n, smatvec_mult, spsolve, b, x, resid, restrt, &iter, stdout);
t = SuperLU_timer_() - t;
/* Output the result. */
nrmA = snrm2_(&(Astore->nnz), (float *)((DNformat *)A.Store)->nzval,
&i_1);
nrmB = snrm2_(&m, b, &i_1);
sp_sgemv("N", -1.0, &A, x, 1, 1.0, b, 1);
res = snrm2_(&m, b, &i_1);
resid = res / nrmB;
printf("||A||_F = %.1e, ||B||_2 = %.1e, ||B-A*X||_2 = %.1e, "
"relres = %.1e\n", nrmA, nrmB, res, resid);
if (iter >= maxit)
{
if (resid >= 1.0) iter = -180;
else if (resid > 1e-8) iter = -111;
}
printf("iteration: %d\nresidual: %.1e\nGMRES time: %.2f seconds.\n",
iter, resid, t);
/* Scale the solution back if equilibration was performed. */
if (*equed == 'C' || *equed == 'B')
for (i = 0; i < n; i++) x[i] *= C[i];
for (i = 0; i < m; i++) {
maxferr = SUPERLU_MAX(maxferr, fabs(x[i] - xact[i]));
}
printf("||X-X_true||_oo = %.1e\n", maxferr);
}
#ifdef DEBUG
printf("%d entries in L and %d entries in U dropped.\n",
num_drop_L, num_drop_U);
#endif
fflush(stdout);
if ( options.PrintStat ) StatPrint(&stat);
StatFree(&stat);
SUPERLU_FREE (rhsb);
SUPERLU_FREE (rhsx);
SUPERLU_FREE (xact);
SUPERLU_FREE (etree);
SUPERLU_FREE (perm_r);
SUPERLU_FREE (perm_c);
SUPERLU_FREE (R);
SUPERLU_FREE (C);
Destroy_CompCol_Matrix(&A);
Destroy_SuperMatrix_Store(&B);
Destroy_SuperMatrix_Store(&X);
if ( lwork >= 0 ) {
Destroy_SuperNode_Matrix(&L);
Destroy_CompCol_Matrix(&U);
}
SUPERLU_FREE(b);
SUPERLU_FREE(x);
#if ( DEBUGlevel>=1 )
CHECK_MALLOC("Exit main()");
#endif
return 0;
}
int main(int argc, char *argv[])
{
char equed[1];
yes_no_t equil;
trans_t trans;
SuperMatrix A, L, U;
SuperMatrix B, X;
NCformat *Astore;
NCformat *Ustore;
SCformat *Lstore;
float *a;
int *asub, *xa;
int *perm_r; /* row permutations from partial pivoting */
int *perm_c; /* column permutation vector */
int *etree;
void *work;
int info, lwork, nrhs, ldx;
int i, m, n, nnz;
float *rhsb, *rhsx, *xact;
float *R, *C;
float *ferr, *berr;
float u, rpg, rcond;
mem_usage_t mem_usage;
superlu_options_t options;
SuperLUStat_t stat;
extern void parse_command_line();
#if ( DEBUGlevel>=1 )
CHECK_MALLOC("Enter main()");
#endif
/* Defaults */
lwork = 0;
nrhs = 1;
equil = YES;
u = 1.0;
trans = NOTRANS;
/* Set the default input options:
options.Fact = DOFACT;
options.Equil = YES;
options.ColPerm = COLAMD;
options.DiagPivotThresh = 1.0;
options.Trans = NOTRANS;
options.IterRefine = NOREFINE;
options.SymmetricMode = NO;
options.PivotGrowth = NO;
options.ConditionNumber = NO;
options.PrintStat = YES;
*/
set_default_options(&options);
/* Can use command line input to modify the defaults. */
parse_command_line(argc, argv, &lwork, &u, &equil, &trans);
options.Equil = equil;
options.DiagPivotThresh = u;
options.Trans = trans;
/* Add more functionalities that the defaults. */
options.PivotGrowth = YES; /* Compute reciprocal pivot growth */
options.ConditionNumber = YES;/* Compute reciprocal condition number */
options.IterRefine = SLU_SINGLE; /* Perform single-precision refinement */
if ( lwork > 0 ) {
work = SUPERLU_MALLOC(lwork);
if ( !work ) {
ABORT("SLINSOLX: cannot allocate work[]");
}
}
/* Read matrix A from a file in Harwell-Boeing format.*/
sreadhb(&m, &n, &nnz, &a, &asub, &xa);
sCreate_CompCol_Matrix(&A, m, n, nnz, a, asub, xa, SLU_NC, SLU_S, SLU_GE);
Astore = A.Store;
printf("Dimension %dx%d; # nonzeros %d\n", A.nrow, A.ncol, Astore->nnz);
if ( !(rhsb = floatMalloc(m * nrhs)) ) ABORT("Malloc fails for rhsb[].");
if ( !(rhsx = floatMalloc(m * nrhs)) ) ABORT("Malloc fails for rhsx[].");
sCreate_Dense_Matrix(&B, m, nrhs, rhsb, m, SLU_DN, SLU_S, SLU_GE);
sCreate_Dense_Matrix(&X, m, nrhs, rhsx, m, SLU_DN, SLU_S, SLU_GE);
xact = floatMalloc(n * nrhs);
ldx = n;
sGenXtrue(n, nrhs, xact, ldx);
sFillRHS(trans, nrhs, xact, ldx, &A, &B);
if ( !(etree = intMalloc(n)) ) ABORT("Malloc fails for etree[].");
if ( !(perm_r = intMalloc(m)) ) ABORT("Malloc fails for perm_r[].");
if ( !(perm_c = intMalloc(n)) ) ABORT("Malloc fails for perm_c[].");
if ( !(R = (float *) SUPERLU_MALLOC(A.nrow * sizeof(float))) )
ABORT("SUPERLU_MALLOC fails for R[].");
if ( !(C = (float *) SUPERLU_MALLOC(A.ncol * sizeof(float))) )
ABORT("SUPERLU_MALLOC fails for C[].");
if ( !(ferr = (float *) SUPERLU_MALLOC(nrhs * sizeof(float))) )
ABORT("SUPERLU_MALLOC fails for ferr[].");
if ( !(berr = (float *) SUPERLU_MALLOC(nrhs * sizeof(float))) )
ABORT("SUPERLU_MALLOC fails for berr[].");
/* Initialize the statistics variables. */
StatInit(&stat);
/* Solve the system and compute the condition number
and error bounds using dgssvx. */
sgssvx(&options, &A, perm_c, perm_r, etree, equed, R, C,
&L, &U, work, lwork, &B, &X, &rpg, &rcond, ferr, berr,
&mem_usage, &stat, &info);
printf("sgssvx(): info %d\n", info);
if ( info == 0 || info == n+1 ) {
/* This is how you could access the solution matrix. */
float *sol = (float*) ((DNformat*) X.Store)->nzval;
if ( options.PivotGrowth == YES )
printf("Recip. pivot growth = %e\n", rpg);
if ( options.ConditionNumber == YES )
printf("Recip. condition number = %e\n", rcond);
if ( options.IterRefine != NOREFINE ) {
printf("Iterative Refinement:\n");
printf("%8s%8s%16s%16s\n", "rhs", "Steps", "FERR", "BERR");
for (i = 0; i < nrhs; ++i)
printf("%8d%8d%16e%16e\n", i+1, stat.RefineSteps, ferr[i], berr[i]);
}
Lstore = (SCformat *) L.Store;
Ustore = (NCformat *) U.Store;
printf("No of nonzeros in factor L = %d\n", Lstore->nnz);
printf("No of nonzeros in factor U = %d\n", Ustore->nnz);
printf("No of nonzeros in L+U = %d\n", Lstore->nnz + Ustore->nnz - n);
printf("FILL ratio = %.1f\n", (float)(Lstore->nnz + Ustore->nnz - n)/nnz);
printf("L\\U MB %.3f\ttotal MB needed %.3f\n",
mem_usage.for_lu/1e6, mem_usage.total_needed/1e6);
fflush(stdout);
} else if ( info > 0 && lwork == -1 ) {
printf("** Estimated memory: %d bytes\n", info - n);
}
if ( options.PrintStat ) StatPrint(&stat);
StatFree(&stat);
SUPERLU_FREE (rhsb);
SUPERLU_FREE (rhsx);
SUPERLU_FREE (xact);
SUPERLU_FREE (etree);
SUPERLU_FREE (perm_r);
SUPERLU_FREE (perm_c);
SUPERLU_FREE (R);
SUPERLU_FREE (C);
SUPERLU_FREE (ferr);
SUPERLU_FREE (berr);
Destroy_CompCol_Matrix(&A);
Destroy_SuperMatrix_Store(&B);
Destroy_SuperMatrix_Store(&X);
if ( lwork == 0 ) {
Destroy_SuperNode_Matrix(&L);
Destroy_CompCol_Matrix(&U);
} else if ( lwork > 0 ) {
SUPERLU_FREE(work);
}
#if ( DEBUGlevel>=1 )
CHECK_MALLOC("Exit main()");
#endif
}
main(int argc, char *argv[])
{
/*
* Purpose
* =======
*
* SDRIVE is the main test program for the FLOAT linear
* equation driver routines SGSSV and SGSSVX.
*
* The program is invoked by a shell script file -- stest.csh.
* The output from the tests are written into a file -- stest.out.
*
* =====================================================================
*/
float *a, *a_save;
int *asub, *asub_save;
int *xa, *xa_save;
SuperMatrix A, B, X, L, U;
SuperMatrix ASAV, AC;
GlobalLU_t Glu; /* Not needed on return. */
mem_usage_t mem_usage;
int *perm_r; /* row permutation from partial pivoting */
int *perm_c, *pc_save; /* column permutation */
int *etree;
float zero = 0.0;
float *R, *C;
float *ferr, *berr;
float *rwork;
float *wwork;
void *work;
int info, lwork, nrhs, panel_size, relax;
int m, n, nnz;
float *xact;
float *rhsb, *solx, *bsav;
int ldb, ldx;
float rpg, rcond;
int i, j, k1;
float rowcnd, colcnd, amax;
int maxsuper, rowblk, colblk;
int prefact, nofact, equil, iequed;
int nt, nrun, nfail, nerrs, imat, fimat, nimat;
int nfact, ifact, itran;
int kl, ku, mode, lda;
int zerot, izero, ioff;
double u;
float anorm, cndnum;
float *Afull;
float result[NTESTS];
superlu_options_t options;
fact_t fact;
trans_t trans;
SuperLUStat_t stat;
static char matrix_type[8];
static char equed[1], path[4], sym[1], dist[1];
FILE *fp;
/* Fixed set of parameters */
int iseed[] = {1988, 1989, 1990, 1991};
static char equeds[] = {'N', 'R', 'C', 'B'};
static fact_t facts[] = {FACTORED, DOFACT, SamePattern,
SamePattern_SameRowPerm};
static trans_t transs[] = {NOTRANS, TRANS, CONJ};
/* Some function prototypes */
extern int sgst01(int, int, SuperMatrix *, SuperMatrix *,
SuperMatrix *, int *, int *, float *);
extern int sgst02(trans_t, int, int, int, SuperMatrix *, float *,
int, float *, int, float *resid);
extern int sgst04(int, int, float *, int,
float *, int, float rcond, float *resid);
extern int sgst07(trans_t, int, int, SuperMatrix *, float *, int,
float *, int, float *, int,
float *, float *, float *);
extern int slatb4_slu(char *, int *, int *, int *, char *, int *, int *,
float *, int *, float *, char *);
extern int slatms_slu(int *, int *, char *, int *, char *, float *d,
int *, float *, float *, int *, int *,
char *, float *, int *, float *, int *);
extern int sp_sconvert(int, int, float *, int, int, int,
float *a, int *, int *, int *);
/* Executable statements */
strcpy(path, "SGE");
nrun = 0;
nfail = 0;
nerrs = 0;
/* Defaults */
lwork = 0;
n = 1;
nrhs = 1;
panel_size = sp_ienv(1);
relax = sp_ienv(2);
u = 1.0;
strcpy(matrix_type, "LA");
parse_command_line(argc, argv, matrix_type, &n,
&panel_size, &relax, &nrhs, &maxsuper,
&rowblk, &colblk, &lwork, &u, &fp);
if ( lwork > 0 ) {
work = SUPERLU_MALLOC(lwork);
if ( !work ) {
fprintf(stderr, "expert: cannot allocate %d bytes\n", lwork);
exit (-1);
}
}
/* Set the default input options. */
set_default_options(&options);
options.DiagPivotThresh = u;
options.PrintStat = NO;
options.PivotGrowth = YES;
options.ConditionNumber = YES;
options.IterRefine = SLU_SINGLE;
if ( strcmp(matrix_type, "LA") == 0 ) {
/* Test LAPACK matrix suite. */
m = n;
lda = SUPERLU_MAX(n, 1);
nnz = n * n; /* upper bound */
fimat = 1;
nimat = NTYPES;
Afull = floatCalloc(lda * n);
sallocateA(n, nnz, &a, &asub, &xa);
} else {
/* Read a sparse matrix */
fimat = nimat = 0;
sreadhb(fp, &m, &n, &nnz, &a, &asub, &xa);
}
sallocateA(n, nnz, &a_save, &asub_save, &xa_save);
rhsb = floatMalloc(m * nrhs);
bsav = floatMalloc(m * nrhs);
solx = floatMalloc(n * nrhs);
ldb = m;
ldx = n;
sCreate_Dense_Matrix(&B, m, nrhs, rhsb, ldb, SLU_DN, SLU_S, SLU_GE);
sCreate_Dense_Matrix(&X, n, nrhs, solx, ldx, SLU_DN, SLU_S, SLU_GE);
xact = floatMalloc(n * nrhs);
etree = intMalloc(n);
perm_r = intMalloc(n);
perm_c = intMalloc(n);
pc_save = intMalloc(n);
R = (float *) SUPERLU_MALLOC(m*sizeof(float));
C = (float *) SUPERLU_MALLOC(n*sizeof(float));
ferr = (float *) SUPERLU_MALLOC(nrhs*sizeof(float));
berr = (float *) SUPERLU_MALLOC(nrhs*sizeof(float));
j = SUPERLU_MAX(m,n) * SUPERLU_MAX(4,nrhs);
rwork = (float *) SUPERLU_MALLOC(j*sizeof(float));
for (i = 0; i < j; ++i) rwork[i] = 0.;
if ( !R ) ABORT("SUPERLU_MALLOC fails for R");
if ( !C ) ABORT("SUPERLU_MALLOC fails for C");
if ( !ferr ) ABORT("SUPERLU_MALLOC fails for ferr");
if ( !berr ) ABORT("SUPERLU_MALLOC fails for berr");
if ( !rwork ) ABORT("SUPERLU_MALLOC fails for rwork");
wwork = floatCalloc( SUPERLU_MAX(m,n) * SUPERLU_MAX(4,nrhs) );
for (i = 0; i < n; ++i) perm_c[i] = pc_save[i] = i;
options.ColPerm = MY_PERMC;
for (imat = fimat; imat <= nimat; ++imat) { /* All matrix types */
if ( imat ) {
/* Skip types 5, 6, or 7 if the matrix size is too small. */
zerot = (imat >= 5 && imat <= 7);
if ( zerot && n < imat-4 )
continue;
/* Set up parameters with SLATB4 and generate a test matrix
with SLATMS. */
slatb4_slu(path, &imat, &n, &n, sym, &kl, &ku, &anorm, &mode,
&cndnum, dist);
slatms_slu(&n, &n, dist, iseed, sym, &rwork[0], &mode, &cndnum,
&anorm, &kl, &ku, "No packing", Afull, &lda,
&wwork[0], &info);
if ( info ) {
printf(FMT3, "SLATMS", info, izero, n, nrhs, imat, nfail);
continue;
}
/* For types 5-7, zero one or more columns of the matrix
to test that INFO is returned correctly. */
if ( zerot ) {
if ( imat == 5 ) izero = 1;
else if ( imat == 6 ) izero = n;
else izero = n / 2 + 1;
ioff = (izero - 1) * lda;
if ( imat < 7 ) {
for (i = 0; i < n; ++i) Afull[ioff + i] = zero;
} else {
for (j = 0; j < n - izero + 1; ++j)
for (i = 0; i < n; ++i)
Afull[ioff + i + j*lda] = zero;
}
} else {
izero = 0;
}
/* Convert to sparse representation. */
sp_sconvert(n, n, Afull, lda, kl, ku, a, asub, xa, &nnz);
} else {
izero = 0;
zerot = 0;
}
sCreate_CompCol_Matrix(&A, m, n, nnz, a, asub, xa, SLU_NC, SLU_S, SLU_GE);
/* Save a copy of matrix A in ASAV */
sCreate_CompCol_Matrix(&ASAV, m, n, nnz, a_save, asub_save, xa_save,
SLU_NC, SLU_S, SLU_GE);
sCopy_CompCol_Matrix(&A, &ASAV);
/* Form exact solution. */
sGenXtrue(n, nrhs, xact, ldx);
StatInit(&stat);
for (iequed = 0; iequed < 4; ++iequed) {
*equed = equeds[iequed];
if (iequed == 0) nfact = 4;
else nfact = 1; /* Only test factored, pre-equilibrated matrix */
for (ifact = 0; ifact < nfact; ++ifact) {
fact = facts[ifact];
options.Fact = fact;
for (equil = 0; equil < 2; ++equil) {
options.Equil = equil;
prefact = ( options.Fact == FACTORED ||
options.Fact == SamePattern_SameRowPerm );
/* Need a first factor */
nofact = (options.Fact != FACTORED); /* Not factored */
/* Restore the matrix A. */
sCopy_CompCol_Matrix(&ASAV, &A);
if ( zerot ) {
if ( prefact ) continue;
} else if ( options.Fact == FACTORED ) {
if ( equil || iequed ) {
/* Compute row and column scale factors to
equilibrate matrix A. */
sgsequ(&A, R, C, &rowcnd, &colcnd, &amax, &info);
/* Force equilibration. */
if ( !info && n > 0 ) {
if ( strncmp(equed, "R", 1)==0 ) {
rowcnd = 0.;
colcnd = 1.;
} else if ( strncmp(equed, "C", 1)==0 ) {
rowcnd = 1.;
colcnd = 0.;
} else if ( strncmp(equed, "B", 1)==0 ) {
rowcnd = 0.;
colcnd = 0.;
}
}
/* Equilibrate the matrix. */
slaqgs(&A, R, C, rowcnd, colcnd, amax, equed);
}
}
if ( prefact ) { /* Need a factor for the first time */
/* Save Fact option. */
fact = options.Fact;
options.Fact = DOFACT;
/* Preorder the matrix, obtain the column etree. */
sp_preorder(&options, &A, perm_c, etree, &AC);
/* Factor the matrix AC. */
sgstrf(&options, &AC, relax, panel_size,
etree, work, lwork, perm_c, perm_r, &L, &U,
&Glu, &stat, &info);
if ( info ) {
printf("** First factor: info %d, equed %c\n",
info, *equed);
if ( lwork == -1 ) {
printf("** Estimated memory: %d bytes\n",
info - n);
exit(0);
}
}
Destroy_CompCol_Permuted(&AC);
/* Restore Fact option. */
options.Fact = fact;
} /* if .. first time factor */
for (itran = 0; itran < NTRAN; ++itran) {
trans = transs[itran];
options.Trans = trans;
/* Restore the matrix A. */
sCopy_CompCol_Matrix(&ASAV, &A);
/* Set the right hand side. */
sFillRHS(trans, nrhs, xact, ldx, &A, &B);
sCopy_Dense_Matrix(m, nrhs, rhsb, ldb, bsav, ldb);
/*----------------
* Test sgssv
*----------------*/
if ( options.Fact == DOFACT && itran == 0) {
/* Not yet factored, and untransposed */
sCopy_Dense_Matrix(m, nrhs, rhsb, ldb, solx, ldx);
sgssv(&options, &A, perm_c, perm_r, &L, &U, &X,
&stat, &info);
if ( info && info != izero ) {
printf(FMT3, "sgssv",
info, izero, n, nrhs, imat, nfail);
} else {
/* Reconstruct matrix from factors and
compute residual. */
sgst01(m, n, &A, &L, &U, perm_c, perm_r,
&result[0]);
nt = 1;
if ( izero == 0 ) {
/* Compute residual of the computed
solution. */
sCopy_Dense_Matrix(m, nrhs, rhsb, ldb,
wwork, ldb);
sgst02(trans, m, n, nrhs, &A, solx,
ldx, wwork,ldb, &result[1]);
nt = 2;
}
/* Print information about the tests that
did not pass the threshold. */
for (i = 0; i < nt; ++i) {
if ( result[i] >= THRESH ) {
printf(FMT1, "sgssv", n, i,
result[i]);
++nfail;
}
}
nrun += nt;
} /* else .. info == 0 */
/* Restore perm_c. */
for (i = 0; i < n; ++i) perm_c[i] = pc_save[i];
if (lwork == 0) {
Destroy_SuperNode_Matrix(&L);
Destroy_CompCol_Matrix(&U);
}
} /* if .. end of testing sgssv */
/*----------------
* Test sgssvx
*----------------*/
/* Equilibrate the matrix if fact = FACTORED and
equed = 'R', 'C', or 'B'. */
if ( options.Fact == FACTORED &&
(equil || iequed) && n > 0 ) {
slaqgs(&A, R, C, rowcnd, colcnd, amax, equed);
}
/* Solve the system and compute the condition number
and error bounds using sgssvx. */
sgssvx(&options, &A, perm_c, perm_r, etree,
equed, R, C, &L, &U, work, lwork, &B, &X, &rpg,
&rcond, ferr, berr, &Glu,
&mem_usage, &stat, &info);
if ( info && info != izero ) {
printf(FMT3, "sgssvx",
info, izero, n, nrhs, imat, nfail);
if ( lwork == -1 ) {
printf("** Estimated memory: %.0f bytes\n",
mem_usage.total_needed);
exit(0);
}
} else {
if ( !prefact ) {
/* Reconstruct matrix from factors and
compute residual. */
sgst01(m, n, &A, &L, &U, perm_c, perm_r,
&result[0]);
k1 = 0;
} else {
k1 = 1;
}
if ( !info ) {
/* Compute residual of the computed solution.*/
sCopy_Dense_Matrix(m, nrhs, bsav, ldb,
wwork, ldb);
sgst02(trans, m, n, nrhs, &ASAV, solx, ldx,
wwork, ldb, &result[1]);
/* Check solution from generated exact
solution. */
sgst04(n, nrhs, solx, ldx, xact, ldx, rcond,
&result[2]);
/* Check the error bounds from iterative
refinement. */
sgst07(trans, n, nrhs, &ASAV, bsav, ldb,
solx, ldx, xact, ldx, ferr, berr,
&result[3]);
/* Print information about the tests that did
not pass the threshold. */
for (i = k1; i < NTESTS; ++i) {
if ( result[i] >= THRESH ) {
printf(FMT2, "sgssvx",
options.Fact, trans, *equed,
n, imat, i, result[i]);
++nfail;
}
}
nrun += NTESTS;
} /* if .. info == 0 */
} /* else .. end of testing sgssvx */
} /* for itran ... */
if ( lwork == 0 ) {
Destroy_SuperNode_Matrix(&L);
Destroy_CompCol_Matrix(&U);
}
} /* for equil ... */
} /* for ifact ... */
} /* for iequed ... */
#if 0
if ( !info ) {
PrintPerf(&L, &U, &mem_usage, rpg, rcond, ferr, berr, equed);
}
#endif
Destroy_SuperMatrix_Store(&A);
Destroy_SuperMatrix_Store(&ASAV);
StatFree(&stat);
} /* for imat ... */
/* Print a summary of the results. */
PrintSumm("SGE", nfail, nrun, nerrs);
if ( strcmp(matrix_type, "LA") == 0 ) SUPERLU_FREE (Afull);
SUPERLU_FREE (rhsb);
SUPERLU_FREE (bsav);
SUPERLU_FREE (solx);
SUPERLU_FREE (xact);
SUPERLU_FREE (etree);
SUPERLU_FREE (perm_r);
SUPERLU_FREE (perm_c);
SUPERLU_FREE (pc_save);
SUPERLU_FREE (R);
SUPERLU_FREE (C);
SUPERLU_FREE (ferr);
SUPERLU_FREE (berr);
SUPERLU_FREE (rwork);
SUPERLU_FREE (wwork);
Destroy_SuperMatrix_Store(&B);
Destroy_SuperMatrix_Store(&X);
#if 0
Destroy_CompCol_Matrix(&A);
Destroy_CompCol_Matrix(&ASAV);
#else
SUPERLU_FREE(a); SUPERLU_FREE(asub); SUPERLU_FREE(xa);
SUPERLU_FREE(a_save); SUPERLU_FREE(asub_save); SUPERLU_FREE(xa_save);
#endif
if ( lwork > 0 ) {
SUPERLU_FREE (work);
Destroy_SuperMatrix_Store(&L);
Destroy_SuperMatrix_Store(&U);
}
return 0;
}
int main ( int argc, char *argv[] )
/**********************************************************************/
/*
Purpose:
SUPER_LU_S3 solves a sparse system read from a file using SGSSVX.
Discussion:
The sparse matrix is stored in a file using the Harwell-Boeing
sparse matrix format. The file should be assigned to the standard
input of this program. For instance, if the matrix is stored
in the file "g10_rua.txt", the execution command might be:
super_lu_s3 < g10_rua.txt
Modified:
25 April 2004
Reference:
James Demmel, John Gilbert, Xiaoye Li,
SuperLU Users's Guide,
Sections 1 and 2.
Local parameters:
SuperMatrix L, the computed L factor.
int *perm_c, the column permutation vector.
int *perm_r, the row permutations from partial pivoting.
SuperMatrix U, the computed U factor.
*/
{
SuperMatrix A;
NCformat *Astore;
float *a;
int *asub;
SuperMatrix B;
float *berr;
float *C;
char equed[1];
yes_no_t equil;
int *etree;
float *ferr;
int i;
int info;
SuperMatrix L;
int ldx;
SCformat *Lstore;
int lwork;
int m;
mem_usage_t mem_usage;
int n;
int nnz;
int nrhs;
superlu_options_t options;
int *perm_c;
int *perm_r;
float *R;
float rcond;
float *rhsb;
float *rhsx;
float rpg;
float *sol;
SuperLUStat_t stat;
trans_t trans;
SuperMatrix U;
float u;
NCformat *Ustore;
void *work;
SuperMatrix X;
int *xa;
float *xact;
/*
Say hello.
*/
printf ( "\n" );
printf ( "SUPER_LU_S3:\n" );
printf ( " Read a sparse matrix A from standard input,\n");
printf ( " stored in Harwell-Boeing Sparse Matrix format.\n" );
printf ( "\n" );
printf ( " Solve a linear system A * X = B using SGSSVX.\n" );
/*
Defaults
*/
lwork = 0;
nrhs = 1;
equil = YES;
u = 1.0;
trans = NOTRANS;
/*
Set the default input options:
options.Fact = DOFACT;
options.Equil = YES;
options.ColPerm = COLAMD;
options.DiagPivotThresh = 1.0;
options.Trans = NOTRANS;
options.IterRefine = NOREFINE;
options.SymmetricMode = NO;
options.PivotGrowth = NO;
options.ConditionNumber = NO;
options.PrintStat = YES;
*/
set_default_options ( &options );
/*
Can use command line input to modify the defaults.
*/
parse_command_line ( argc, argv, &lwork, &u, &equil, &trans );
options.Equil = equil;
options.DiagPivotThresh = u;
options.Trans = trans;
printf ( "\n" );
printf ( " Length of work array LWORK = %d\n", lwork );
printf ( " Equilibration option EQUIL = %d\n", equil );
printf ( " Diagonal pivot threshhold value U = %f\n", u );
printf ( " Tranpose option TRANS = %d\n", trans );
/*
Add more functionalities that the defaults.
Compute reciprocal pivot growth
*/
options.PivotGrowth = YES;
/*
Compute reciprocal condition number
*/
options.ConditionNumber = YES;
/*
Perform single-precision refinement
*/
options.IterRefine = SINGLE;
if ( 0 < lwork )
{
work = SUPERLU_MALLOC(lwork);
if ( !work )
{
ABORT ( "SUPERLU_MALLOC cannot allocate work[]" );
}
}
/*
Read matrix A from a file in Harwell-Boeing format.
*/
sreadhb ( &m, &n, &nnz, &a, &asub, &xa );
/*
Create storage for a compressed column matrix.
*/
sCreate_CompCol_Matrix ( &A, m, n, nnz, a, asub, xa, SLU_NC, SLU_S, SLU_GE );
Astore = A.Store;
printf ( "\n" );
printf ( " Dimension %dx%d; # nonzeros %d\n", A.nrow, A.ncol, Astore->nnz );
rhsb = floatMalloc ( m * nrhs );
if ( !rhsb )
{
ABORT ( "Malloc fails for rhsb[]." );
}
rhsx = floatMalloc ( m * nrhs );
if ( !rhsx )
{
ABORT ( "Malloc fails for rhsx[]." );
}
sCreate_Dense_Matrix ( &B, m, nrhs, rhsb, m, SLU_DN, SLU_S, SLU_GE );
sCreate_Dense_Matrix ( &X, m, nrhs, rhsx, m, SLU_DN, SLU_S, SLU_GE );
xact = floatMalloc ( n * nrhs );
if ( !xact )
{
ABORT ( "SUPERLU_MALLOC cannot allocate xact[]" );
}
ldx = n;
sGenXtrue ( n, nrhs, xact, ldx );
sFillRHS ( trans, nrhs, xact, ldx, &A, &B );
etree = intMalloc ( n );
if ( !etree )
{
ABORT ( "Malloc fails for etree[]." );
}
perm_c = intMalloc ( n );
if ( !perm_c )
{
ABORT ( "Malloc fails for perm_c[]." );
}
perm_r = intMalloc ( m );
if ( !perm_r )
{
ABORT ( "Malloc fails for perm_r[]." );
}
R = (float *) SUPERLU_MALLOC ( A.nrow * sizeof(float) );
if ( !R )
{
ABORT ( "SUPERLU_MALLOC fails for R[]." );
}
C = (float *) SUPERLU_MALLOC ( A.ncol * sizeof(float) );
if ( !C )
{
ABORT ( "SUPERLU_MALLOC fails for C[]." );
}
ferr = (float *) SUPERLU_MALLOC ( nrhs * sizeof(float) );
if ( !ferr )
{
ABORT ( "SUPERLU_MALLOC fails for ferr[]." );
}
berr = (float *) SUPERLU_MALLOC ( nrhs * sizeof(float) );
if ( !berr )
{
ABORT ( "SUPERLU_MALLOC fails for berr[]." );
}
/*
Initialize the statistics variables.
*/
StatInit(&stat);
/*
Solve the system and compute the condition number and error bounds using SGSSVX.
*/
sgssvx ( &options, &A, perm_c, perm_r, etree, equed, R, C,
&L, &U, work, lwork, &B, &X, &rpg, &rcond, ferr, berr,
&mem_usage, &stat, &info );
printf ( "\n" );
printf ( " SGSSVX returns INFO = %d\n", info );
if ( info == 0 || info == n+1 )
{
sol = (float*) ((DNformat*) X.Store)->nzval;
if ( options.PivotGrowth == YES )
{
printf ( "\n" );
printf ( " Reciprocal pivot growth = %e\n", rpg);
}
if ( options.ConditionNumber == YES )
{
printf ( "\n" );
printf ( " Reciprocal condition number = %e\n", rcond);
}
if ( options.IterRefine != NOREFINE )
{
printf ( "\n" );
printf ( " Iterative Refinement:\n");
printf ( "%8s%8s%16s%16s\n", "rhs", "Steps", "FERR", "BERR");
for ( i = 0; i < nrhs; i++ )
{
printf ( "%8d%8d%16e%16e\n", i+1, stat.RefineSteps, ferr[i], berr[i]);
}
}
Lstore = (SCformat *) L.Store;
Ustore = (NCformat *) U.Store;
printf ( "\n" );
printf ( " Number of nonzeros in factor L = %d\n", Lstore->nnz );
printf ( " Number of nonzeros in factor U = %d\n", Ustore->nnz );
printf ( " Number of nonzeros in L+U = %d\n",
Lstore->nnz + Ustore->nnz - n );
printf ( "\n" );
printf ( " L\\U MB %.3f\ttotal MB needed %.3f\texpansions %d\n",
mem_usage.for_lu/1e6, mem_usage.total_needed/1e6,
mem_usage.expansions );
fflush ( stdout );
}
else if ( info > 0 && lwork == -1 )
{
printf ( "\n" );
printf ( " Estimated memory: %d bytes\n", info - n );
}
if ( options.PrintStat )
{
StatPrint ( &stat );
}
StatFree ( &stat );
SUPERLU_FREE ( rhsb );
SUPERLU_FREE ( rhsx );
SUPERLU_FREE ( xact );
SUPERLU_FREE ( etree );
SUPERLU_FREE ( perm_r );
SUPERLU_FREE ( perm_c );
SUPERLU_FREE ( R );
SUPERLU_FREE ( C );
SUPERLU_FREE ( ferr );
SUPERLU_FREE ( berr );
Destroy_CompCol_Matrix ( &A );
Destroy_SuperMatrix_Store ( &B );
Destroy_SuperMatrix_Store ( &X );
if ( 0 <= lwork )
{
Destroy_SuperNode_Matrix ( &L );
Destroy_CompCol_Matrix ( &U );
}
/*
Say goodbye.
*/
printf ( "\n" );
printf ( "SUPER_LU_S3:\n" );
printf ( " Normal end of execution.\n");
return 0;
}
int main ( int argc, char *argv[] )
/******************************************************************************/
/*
Purpose:
MAIN is the main program for PSLINSOL.
Licensing:
This code is distributed under the GNU LGPL license.
Modified:
10 February 2014
Author:
Xiaoye Li
*/
{
SuperMatrix A;
NCformat *Astore;
float *a;
int *asub, *xa;
int *perm_r; /* row permutations from partial pivoting */
int *perm_c; /* column permutation vector */
SuperMatrix L; /* factor L */
SCPformat *Lstore;
SuperMatrix U; /* factor U */
NCPformat *Ustore;
SuperMatrix B;
int nrhs, ldx, info, m, n, nnz, b;
int nprocs; /* maximum number of processors to use. */
int panel_size, relax, maxsup;
int permc_spec;
trans_t trans;
float *xact, *rhs;
superlu_memusage_t superlu_memusage;
void parse_command_line();
timestamp ( );
printf ( "\n" );
printf ( "PSLINSOL:\n" );
printf ( " C/OpenMP version\n" );
printf ( " Call the OpenMP version of SuperLU to solve a linear system.\n" );
nrhs = 1;
trans = NOTRANS;
nprocs = 1;
n = 1000;
b = 1;
panel_size = sp_ienv(1);
relax = sp_ienv(2);
maxsup = sp_ienv(3);
/*
Check for any commandline input.
*/
parse_command_line ( argc, argv, &nprocs, &n, &b, &panel_size,
&relax, &maxsup );
#if ( PRNTlevel>=1 || DEBUGlevel>=1 )
cpp_defs();
#endif
#define HB
#if defined( DEN )
m = n;
nnz = n * n;
sband(n, n, nnz, &a, &asub, &xa);
#elif defined( BAND )
m = n;
nnz = (2*b+1) * n;
sband(n, b, nnz, &a, &asub, &xa);
#elif defined( BD )
nb = 5;
bs = 200;
m = n = bs * nb;
nnz = bs * bs * nb;
sblockdiag(nb, bs, nnz, &a, &asub, &xa);
#elif defined( HB )
sreadhb(&m, &n, &nnz, &a, &asub, &xa);
#else
sreadmt(&m, &n, &nnz, &a, &asub, &xa);
#endif
sCreate_CompCol_Matrix(&A, m, n, nnz, a, asub, xa, SLU_NC, SLU_S, SLU_GE);
Astore = A.Store;
printf("Dimension %dx%d; # nonzeros %d\n", A.nrow, A.ncol, Astore->nnz);
if (!(rhs = floatMalloc(m * nrhs))) SUPERLU_ABORT("Malloc fails for rhs[].");
sCreate_Dense_Matrix(&B, m, nrhs, rhs, m, SLU_DN, SLU_S, SLU_GE);
xact = floatMalloc(n * nrhs);
ldx = n;
sGenXtrue(n, nrhs, xact, ldx);
sFillRHS(trans, nrhs, xact, ldx, &A, &B);
if (!(perm_r = intMalloc(m))) SUPERLU_ABORT("Malloc fails for perm_r[].");
if (!(perm_c = intMalloc(n))) SUPERLU_ABORT("Malloc fails for perm_c[].");
/*
* Get column permutation vector perm_c[], according to permc_spec:
* permc_spec = 0: natural ordering
* permc_spec = 1: minimum degree ordering on structure of A'*A
* permc_spec = 2: minimum degree ordering on structure of A'+A
* permc_spec = 3: approximate minimum degree for unsymmetric matrices
*/
permc_spec = 1;
get_perm_c(permc_spec, &A, perm_c);
psgssv(nprocs, &A, perm_c, perm_r, &L, &U, &B, &info);
if ( info == 0 ) {
sinf_norm_error(nrhs, &B, xact); /* Inf. norm of the error */
Lstore = (SCPformat *) L.Store;
Ustore = (NCPformat *) U.Store;
printf("#NZ in factor L = %d\n", Lstore->nnz);
printf("#NZ in factor U = %d\n", Ustore->nnz);
printf("#NZ in L+U = %d\n", Lstore->nnz + Ustore->nnz - L.ncol);
superlu_sQuerySpace(nprocs, &L, &U, panel_size, &superlu_memusage);
printf("L\\U MB %.3f\ttotal MB needed %.3f\texpansions %d\n",
superlu_memusage.for_lu/1024/1024,
superlu_memusage.total_needed/1024/1024,
superlu_memusage.expansions);
}
SUPERLU_FREE (rhs);
SUPERLU_FREE (xact);
SUPERLU_FREE (perm_r);
SUPERLU_FREE (perm_c);
Destroy_CompCol_Matrix(&A);
Destroy_SuperMatrix_Store(&B);
Destroy_SuperNode_SCP(&L);
Destroy_CompCol_NCP(&U);
/*
Terminate.
*/
printf ( "\n" );
printf ( "PSLINSOL:\n" );
printf ( " Normal end of execution.\n" );
printf ( "\n" );
timestamp ( );
return 0;
}
main(int argc, char *argv[])
{
/*
* Purpose
* =======
*
* The driver program SLINSOLX1.
*
* This example illustrates how to use SGSSVX to solve systems with the same
* A but different right-hand side.
* In this case, we factorize A only once in the first call to DGSSVX,
* and reuse the following data structures in the subsequent call to SGSSVX:
* perm_c, perm_r, R, C, L, U.
*
*/
char equed[1];
yes_no_t equil;
trans_t trans;
SuperMatrix A, L, U;
SuperMatrix B, X;
NCformat *Astore;
NCformat *Ustore;
SCformat *Lstore;
float *a;
int *asub, *xa;
int *perm_c; /* column permutation vector */
int *perm_r; /* row permutations from partial pivoting */
int *etree;
void *work;
int info, lwork, nrhs, ldx;
int i, m, n, nnz;
float *rhsb, *rhsx, *xact;
float *R, *C;
float *ferr, *berr;
float u, rpg, rcond;
mem_usage_t mem_usage;
superlu_options_t options;
SuperLUStat_t stat;
extern void parse_command_line();
#if ( DEBUGlevel>=1 )
CHECK_MALLOC("Enter main()");
#endif
/* Defaults */
lwork = 0;
nrhs = 1;
equil = YES;
u = 1.0;
trans = NOTRANS;
/* Set the default values for options argument:
options.Fact = DOFACT;
options.Equil = YES;
options.ColPerm = COLAMD;
options.DiagPivotThresh = 1.0;
options.Trans = NOTRANS;
options.IterRefine = NOREFINE;
options.SymmetricMode = NO;
options.PivotGrowth = NO;
options.ConditionNumber = NO;
options.PrintStat = YES;
*/
set_default_options(&options);
/* Can use command line input to modify the defaults. */
parse_command_line(argc, argv, &lwork, &u, &equil, &trans);
options.Equil = equil;
options.DiagPivotThresh = u;
options.Trans = trans;
if ( lwork > 0 ) {
work = SUPERLU_MALLOC(lwork);
if ( !work ) {
ABORT("SLINSOLX: cannot allocate work[]");
}
}
/* Read matrix A from a file in Harwell-Boeing format.*/
sreadhb(&m, &n, &nnz, &a, &asub, &xa);
sCreate_CompCol_Matrix(&A, m, n, nnz, a, asub, xa, SLU_NC, SLU_S, SLU_GE);
Astore = A.Store;
printf("Dimension %dx%d; # nonzeros %d\n", A.nrow, A.ncol, Astore->nnz);
if ( !(rhsb = floatMalloc(m * nrhs)) ) ABORT("Malloc fails for rhsb[].");
if ( !(rhsx = floatMalloc(m * nrhs)) ) ABORT("Malloc fails for rhsx[].");
sCreate_Dense_Matrix(&B, m, nrhs, rhsb, m, SLU_DN, SLU_S, SLU_GE);
sCreate_Dense_Matrix(&X, m, nrhs, rhsx, m, SLU_DN, SLU_S, SLU_GE);
xact = floatMalloc(n * nrhs);
ldx = n;
sGenXtrue(n, nrhs, xact, ldx);
sFillRHS(trans, nrhs, xact, ldx, &A, &B);
if ( !(etree = intMalloc(n)) ) ABORT("Malloc fails for etree[].");
if ( !(perm_r = intMalloc(m)) ) ABORT("Malloc fails for perm_r[].");
if ( !(perm_c = intMalloc(n)) ) ABORT("Malloc fails for perm_c[].");
if ( !(R = (float *) SUPERLU_MALLOC(A.nrow * sizeof(float))) )
ABORT("SUPERLU_MALLOC fails for R[].");
if ( !(C = (float *) SUPERLU_MALLOC(A.ncol * sizeof(float))) )
ABORT("SUPERLU_MALLOC fails for C[].");
if ( !(ferr = (float *) SUPERLU_MALLOC(nrhs * sizeof(float))) )
ABORT("SUPERLU_MALLOC fails for ferr[].");
if ( !(berr = (float *) SUPERLU_MALLOC(nrhs * sizeof(float))) )
ABORT("SUPERLU_MALLOC fails for berr[].");
/* Initialize the statistics variables. */
StatInit(&stat);
/* ONLY PERFORM THE LU DECOMPOSITION */
B.ncol = 0; /* Indicate not to solve the system */
sgssvx(&options, &A, perm_c, perm_r, etree, equed, R, C,
&L, &U, work, lwork, &B, &X, &rpg, &rcond, ferr, berr,
&mem_usage, &stat, &info);
printf("LU factorization: sgssvx() returns info %d\n", info);
if ( info == 0 || info == n+1 ) {
if ( options.PivotGrowth ) printf("Recip. pivot growth = %e\n", rpg);
if ( options.ConditionNumber )
printf("Recip. condition number = %e\n", rcond);
Lstore = (SCformat *) L.Store;
Ustore = (NCformat *) U.Store;
printf("No of nonzeros in factor L = %d\n", Lstore->nnz);
printf("No of nonzeros in factor U = %d\n", Ustore->nnz);
printf("No of nonzeros in L+U = %d\n", Lstore->nnz + Ustore->nnz - n);
printf("FILL ratio = %.1f\n", (float)(Lstore->nnz + Ustore->nnz - n)/nnz);
printf("L\\U MB %.3f\ttotal MB needed %.3f\n",
mem_usage.for_lu/1e6, mem_usage.total_needed/1e6);
fflush(stdout);
} else if ( info > 0 && lwork == -1 ) {
printf("** Estimated memory: %d bytes\n", info - n);
}
if ( options.PrintStat ) StatPrint(&stat);
StatFree(&stat);
/* ------------------------------------------------------------
NOW WE SOLVE THE LINEAR SYSTEM USING THE FACTORED FORM OF A.
------------------------------------------------------------*/
options.Fact = FACTORED; /* Indicate the factored form of A is supplied. */
B.ncol = nrhs; /* Set the number of right-hand side */
/* Initialize the statistics variables. */
StatInit(&stat);
sgssvx(&options, &A, perm_c, perm_r, etree, equed, R, C,
&L, &U, work, lwork, &B, &X, &rpg, &rcond, ferr, berr,
&mem_usage, &stat, &info);
printf("Triangular solve: sgssvx() returns info %d\n", info);
if ( info == 0 || info == n+1 ) {
/* This is how you could access the solution matrix. */
float *sol = (float*) ((DNformat*) X.Store)->nzval;
if ( options.IterRefine ) {
printf("Iterative Refinement:\n");
printf("%8s%8s%16s%16s\n", "rhs", "Steps", "FERR", "BERR");
for (i = 0; i < nrhs; ++i)
printf("%8d%8d%16e%16e\n", i+1, stat.RefineSteps, ferr[i], berr[i]);
}
fflush(stdout);
} else if ( info > 0 && lwork == -1 ) {
printf("** Estimated memory: %d bytes\n", info - n);
}
if ( options.PrintStat ) StatPrint(&stat);
StatFree(&stat);
SUPERLU_FREE (rhsb);
SUPERLU_FREE (rhsx);
SUPERLU_FREE (xact);
SUPERLU_FREE (etree);
SUPERLU_FREE (perm_r);
SUPERLU_FREE (perm_c);
SUPERLU_FREE (R);
SUPERLU_FREE (C);
SUPERLU_FREE (ferr);
SUPERLU_FREE (berr);
Destroy_CompCol_Matrix(&A);
Destroy_SuperMatrix_Store(&B);
Destroy_SuperMatrix_Store(&X);
if ( lwork >= 0 ) {
Destroy_SuperNode_Matrix(&L);
Destroy_CompCol_Matrix(&U);
}
#if ( DEBUGlevel>=1 )
CHECK_MALLOC("Exit main()");
#endif
}
main(int argc, char *argv[])
{
/*
* Purpose
* =======
*
* The driver program SLINSOLX2.
*
* This example illustrates how to use SGSSVX to solve systems repeatedly
* with the same sparsity pattern of matrix A.
* In this case, the column permutation vector perm_c is computed once.
* The following data structures will be reused in the subsequent call to
* SGSSVX: perm_c, etree
*
*/
char equed[1];
yes_no_t equil;
trans_t trans;
SuperMatrix A, A1, L, U;
SuperMatrix B, B1, X;
NCformat *Astore;
NCformat *Ustore;
SCformat *Lstore;
float *a, *a1;
int *asub, *xa, *asub1, *xa1;
int *perm_r; /* row permutations from partial pivoting */
int *perm_c; /* column permutation vector */
int *etree;
void *work;
int info, lwork, nrhs, ldx;
int i, j, m, n, nnz;
float *rhsb, *rhsb1, *rhsx, *xact;
float *R, *C;
float *ferr, *berr;
float u, rpg, rcond;
mem_usage_t mem_usage;
superlu_options_t options;
SuperLUStat_t stat;
extern void parse_command_line();
#if ( DEBUGlevel>=1 )
CHECK_MALLOC("Enter main()");
#endif
/* Defaults */
lwork = 0;
nrhs = 1;
equil = YES;
u = 1.0;
trans = NOTRANS;
/* Set the default input options:
options.Fact = DOFACT;
options.Equil = YES;
options.ColPerm = COLAMD;
options.DiagPivotThresh = 1.0;
options.Trans = NOTRANS;
options.IterRefine = NOREFINE;
options.SymmetricMode = NO;
options.PivotGrowth = NO;
options.ConditionNumber = NO;
options.PrintStat = YES;
*/
set_default_options(&options);
/* Can use command line input to modify the defaults. */
parse_command_line(argc, argv, &lwork, &u, &equil, &trans);
options.Equil = equil;
options.DiagPivotThresh = u;
options.Trans = trans;
if ( lwork > 0 ) {
work = SUPERLU_MALLOC(lwork);
if ( !work ) {
ABORT("DLINSOLX: cannot allocate work[]");
}
}
/* Read matrix A from a file in Harwell-Boeing format.*/
sreadhb(&m, &n, &nnz, &a, &asub, &xa);
if ( !(a1 = floatMalloc(nnz)) ) ABORT("Malloc fails for a1[].");
if ( !(asub1 = intMalloc(nnz)) ) ABORT("Malloc fails for asub1[].");
if ( !(xa1 = intMalloc(n+1)) ) ABORT("Malloc fails for xa1[].");
for (i = 0; i < nnz; ++i) {
a1[i] = a[i];
asub1[i] = asub[i];
}
for (i = 0; i < n+1; ++i) xa1[i] = xa[i];
sCreate_CompCol_Matrix(&A, m, n, nnz, a, asub, xa, SLU_NC, SLU_S, SLU_GE);
Astore = A.Store;
printf("Dimension %dx%d; # nonzeros %d\n", A.nrow, A.ncol, Astore->nnz);
if ( !(rhsb = floatMalloc(m * nrhs)) ) ABORT("Malloc fails for rhsb[].");
if ( !(rhsb1 = floatMalloc(m * nrhs)) ) ABORT("Malloc fails for rhsb1[].");
if ( !(rhsx = floatMalloc(m * nrhs)) ) ABORT("Malloc fails for rhsx[].");
sCreate_Dense_Matrix(&B, m, nrhs, rhsb, m, SLU_DN, SLU_S, SLU_GE);
sCreate_Dense_Matrix(&X, m, nrhs, rhsx, m, SLU_DN, SLU_S, SLU_GE);
xact = floatMalloc(n * nrhs);
ldx = n;
sGenXtrue(n, nrhs, xact, ldx);
sFillRHS(trans, nrhs, xact, ldx, &A, &B);
for (j = 0; j < nrhs; ++j)
for (i = 0; i < m; ++i) rhsb1[i+j*m] = rhsb[i+j*m];
if ( !(perm_c = intMalloc(n)) ) ABORT("Malloc fails for perm_c[].");
if ( !(perm_r = intMalloc(m)) ) ABORT("Malloc fails for perm_r[].");
if ( !(etree = intMalloc(n)) ) ABORT("Malloc fails for etree[].");
if ( !(R = (float *) SUPERLU_MALLOC(A.nrow * sizeof(float))) )
ABORT("SUPERLU_MALLOC fails for R[].");
if ( !(C = (float *) SUPERLU_MALLOC(A.ncol * sizeof(float))) )
ABORT("SUPERLU_MALLOC fails for C[].");
if ( !(ferr = (float *) SUPERLU_MALLOC(nrhs * sizeof(float))) )
ABORT("SUPERLU_MALLOC fails for ferr[].");
if ( !(berr = (float *) SUPERLU_MALLOC(nrhs * sizeof(float))) )
ABORT("SUPERLU_MALLOC fails for berr[].");
/* Initialize the statistics variables. */
StatInit(&stat);
/* ------------------------------------------------------------
WE SOLVE THE LINEAR SYSTEM FOR THE FIRST TIME: AX = B
------------------------------------------------------------*/
sgssvx(&options, &A, perm_c, perm_r, etree, equed, R, C,
&L, &U, work, lwork, &B, &X, &rpg, &rcond, ferr, berr,
&mem_usage, &stat, &info);
printf("First system: sgssvx() returns info %d\n", info);
if ( info == 0 || info == n+1 ) {
/* This is how you could access the solution matrix. */
float *sol = (float*) ((DNformat*) X.Store)->nzval;
if ( options.PivotGrowth ) printf("Recip. pivot growth = %e\n", rpg);
if ( options.ConditionNumber )
printf("Recip. condition number = %e\n", rcond);
Lstore = (SCformat *) L.Store;
Ustore = (NCformat *) U.Store;
printf("No of nonzeros in factor L = %d\n", Lstore->nnz);
printf("No of nonzeros in factor U = %d\n", Ustore->nnz);
printf("No of nonzeros in L+U = %d\n", Lstore->nnz + Ustore->nnz - n);
printf("FILL ratio = %.1f\n", (float)(Lstore->nnz + Ustore->nnz - n)/nnz);
printf("L\\U MB %.3f\ttotal MB needed %.3f\n",
mem_usage.for_lu/1e6, mem_usage.total_needed/1e6);
if ( options.IterRefine ) {
printf("Iterative Refinement:\n");
printf("%8s%8s%16s%16s\n", "rhs", "Steps", "FERR", "BERR");
for (i = 0; i < nrhs; ++i)
printf("%8d%8d%16e%16e\n", i+1, stat.RefineSteps, ferr[i], berr[i]);
}
fflush(stdout);
} else if ( info > 0 && lwork == -1 ) {
printf("** Estimated memory: %d bytes\n", info - n);
}
if ( options.PrintStat ) StatPrint(&stat);
StatFree(&stat);
Destroy_CompCol_Matrix(&A);
Destroy_Dense_Matrix(&B);
if ( lwork >= 0 ) { /* Deallocate storage associated with L and U. */
Destroy_SuperNode_Matrix(&L);
Destroy_CompCol_Matrix(&U);
}
/* ------------------------------------------------------------
NOW WE SOLVE ANOTHER LINEAR SYSTEM: A1*X = B1
ONLY THE SPARSITY PATTERN OF A1 IS THE SAME AS THAT OF A.
------------------------------------------------------------*/
options.Fact = SamePattern;
StatInit(&stat); /* Initialize the statistics variables. */
sCreate_CompCol_Matrix(&A1, m, n, nnz, a1, asub1, xa1,
SLU_NC, SLU_S, SLU_GE);
sCreate_Dense_Matrix(&B1, m, nrhs, rhsb1, m, SLU_DN, SLU_S, SLU_GE);
sgssvx(&options, &A1, perm_c, perm_r, etree, equed, R, C,
&L, &U, work, lwork, &B1, &X, &rpg, &rcond, ferr, berr,
&mem_usage, &stat, &info);
printf("\nSecond system: sgssvx() returns info %d\n", info);
if ( info == 0 || info == n+1 ) {
/* This is how you could access the solution matrix. */
float *sol = (float*) ((DNformat*) X.Store)->nzval;
if ( options.PivotGrowth ) printf("Recip. pivot growth = %e\n", rpg);
if ( options.ConditionNumber )
printf("Recip. condition number = %e\n", rcond);
Lstore = (SCformat *) L.Store;
Ustore = (NCformat *) U.Store;
printf("No of nonzeros in factor L = %d\n", Lstore->nnz);
printf("No of nonzeros in factor U = %d\n", Ustore->nnz);
printf("No of nonzeros in L+U = %d\n", Lstore->nnz + Ustore->nnz - n);
printf("L\\U MB %.3f\ttotal MB needed %.3f\n",
mem_usage.for_lu/1e6, mem_usage.total_needed/1e6);
if ( options.IterRefine ) {
printf("Iterative Refinement:\n");
printf("%8s%8s%16s%16s\n", "rhs", "Steps", "FERR", "BERR");
for (i = 0; i < nrhs; ++i)
printf("%8d%8d%16e%16e\n", i+1, stat.RefineSteps, ferr[i], berr[i]);
}
fflush(stdout);
} else if ( info > 0 && lwork == -1 ) {
printf("** Estimated memory: %d bytes\n", info - n);
}
if ( options.PrintStat ) StatPrint(&stat);
StatFree(&stat);
SUPERLU_FREE (xact);
SUPERLU_FREE (etree);
SUPERLU_FREE (perm_r);
SUPERLU_FREE (perm_c);
SUPERLU_FREE (R);
SUPERLU_FREE (C);
SUPERLU_FREE (ferr);
SUPERLU_FREE (berr);
Destroy_CompCol_Matrix(&A1);
Destroy_Dense_Matrix(&B1);
Destroy_Dense_Matrix(&X);
if ( lwork == 0 ) {
Destroy_SuperNode_Matrix(&L);
Destroy_CompCol_Matrix(&U);
} else if ( lwork > 0 ) {
SUPERLU_FREE(work);
}
#if ( DEBUGlevel>=1 )
CHECK_MALLOC("Exit main()");
#endif
}
int main(int argc, char *argv[])
{
SuperMatrix A;
NCformat *Astore;
float *a;
int *asub, *xa;
int *perm_c; /* column permutation vector */
int *perm_r; /* row permutations from partial pivoting */
SuperMatrix L; /* factor L */
SCformat *Lstore;
SuperMatrix U; /* factor U */
NCformat *Ustore;
SuperMatrix B;
int nrhs, ldx, info, m, n, nnz;
float *xact, *rhs;
mem_usage_t mem_usage;
superlu_options_t options;
SuperLUStat_t stat;
FILE *fp = stdin;
#if ( DEBUGlevel>=1 )
CHECK_MALLOC("Enter main()");
#endif
/* Set the default input options:
options.Fact = DOFACT;
options.Equil = YES;
options.ColPerm = COLAMD;
options.DiagPivotThresh = 1.0;
options.Trans = NOTRANS;
options.IterRefine = NOREFINE;
options.SymmetricMode = NO;
options.PivotGrowth = NO;
options.ConditionNumber = NO;
options.PrintStat = YES;
*/
set_default_options(&options);
/* Read the matrix in Harwell-Boeing format. */
sreadhb(fp, &m, &n, &nnz, &a, &asub, &xa);
sCreate_CompCol_Matrix(&A, m, n, nnz, a, asub, xa, SLU_NC, SLU_S, SLU_GE);
Astore = A.Store;
printf("Dimension %dx%d; # nonzeros %d\n", A.nrow, A.ncol, Astore->nnz);
nrhs = 1;
if ( !(rhs = floatMalloc(m * nrhs)) ) ABORT("Malloc fails for rhs[].");
sCreate_Dense_Matrix(&B, m, nrhs, rhs, m, SLU_DN, SLU_S, SLU_GE);
xact = floatMalloc(n * nrhs);
ldx = n;
sGenXtrue(n, nrhs, xact, ldx);
sFillRHS(options.Trans, nrhs, xact, ldx, &A, &B);
if ( !(perm_c = intMalloc(n)) ) ABORT("Malloc fails for perm_c[].");
if ( !(perm_r = intMalloc(m)) ) ABORT("Malloc fails for perm_r[].");
/* Initialize the statistics variables. */
StatInit(&stat);
sgssv(&options, &A, perm_c, perm_r, &L, &U, &B, &stat, &info);
if ( info == 0 ) {
/* This is how you could access the solution matrix. */
float *sol = (float*) ((DNformat*) B.Store)->nzval;
/* Compute the infinity norm of the error. */
sinf_norm_error(nrhs, &B, xact);
Lstore = (SCformat *) L.Store;
Ustore = (NCformat *) U.Store;
printf("No of nonzeros in factor L = %d\n", Lstore->nnz);
printf("No of nonzeros in factor U = %d\n", Ustore->nnz);
printf("No of nonzeros in L+U = %d\n", Lstore->nnz + Ustore->nnz - n);
printf("FILL ratio = %.1f\n", (float)(Lstore->nnz + Ustore->nnz - n)/nnz);
sQuerySpace(&L, &U, &mem_usage);
printf("L\\U MB %.3f\ttotal MB needed %.3f\n",
mem_usage.for_lu/1e6, mem_usage.total_needed/1e6);
} else {
printf("sgssv() error returns INFO= %d\n", info);
if ( info <= n ) { /* factorization completes */
sQuerySpace(&L, &U, &mem_usage);
printf("L\\U MB %.3f\ttotal MB needed %.3f\n",
mem_usage.for_lu/1e6, mem_usage.total_needed/1e6);
}
}
if ( options.PrintStat ) StatPrint(&stat);
StatFree(&stat);
SUPERLU_FREE (rhs);
SUPERLU_FREE (xact);
SUPERLU_FREE (perm_r);
SUPERLU_FREE (perm_c);
Destroy_CompCol_Matrix(&A);
Destroy_SuperMatrix_Store(&B);
Destroy_SuperNode_Matrix(&L);
Destroy_CompCol_Matrix(&U);
#if ( DEBUGlevel>=1 )
CHECK_MALLOC("Exit main()");
#endif
}
/* Here is a driver inspired by A. Sheffer's "cow flattener". */
static NLboolean __nlSolve_SUPERLU( NLboolean do_perm) {
/* OpenNL Context */
__NLSparseMatrix* M = &(__nlCurrentContext->M);
NLfloat* b = __nlCurrentContext->b;
NLfloat* x = __nlCurrentContext->x;
/* Compressed Row Storage matrix representation */
NLuint n = __nlCurrentContext->n;
NLuint nnz = __nlSparseMatrixNNZ(M); /* Number of Non-Zero coeffs */
NLint* xa = __NL_NEW_ARRAY(NLint, n+1);
NLfloat* rhs = __NL_NEW_ARRAY(NLfloat, n);
NLfloat* a = __NL_NEW_ARRAY(NLfloat, nnz);
NLint* asub = __NL_NEW_ARRAY(NLint, nnz);
/* Permutation vector */
NLint* perm_r = __NL_NEW_ARRAY(NLint, n);
NLint* perm = __NL_NEW_ARRAY(NLint, n);
/* SuperLU variables */
SuperMatrix A, B; /* System */
SuperMatrix L, U; /* Inverse of A */
NLint info; /* status code */
DNformat *vals = NULL; /* access to result */
float *rvals = NULL; /* access to result */
/* SuperLU options and stats */
superlu_options_t options;
SuperLUStat_t stat;
/* Temporary variables */
__NLRowColumn* Ri = NULL;
NLuint i,jj,count;
__nl_assert(!(M->storage & __NL_SYMMETRIC));
__nl_assert(M->storage & __NL_ROWS);
__nl_assert(M->m == M->n);
/*
* Step 1: convert matrix M into SuperLU compressed column
* representation.
* -------------------------------------------------------
*/
count = 0;
for(i=0; i<n; i++) {
Ri = &(M->row[i]);
xa[i] = count;
for(jj=0; jj<Ri->size; jj++) {
a[count] = Ri->coeff[jj].value;
asub[count] = Ri->coeff[jj].index;
count++;
}
}
xa[n] = nnz;
/* Save memory for SuperLU */
__nlSparseMatrixClear(M);
/*
* Rem: symmetric storage does not seem to work with
* SuperLU ... (->deactivated in main SLS::Solver driver)
*/
sCreate_CompCol_Matrix(
&A, n, n, nnz, a, asub, xa,
SLU_NR, /* Row_wise, no supernode */
SLU_S, /* floats */
SLU_GE /* general storage */
);
/* Step 2: create vector */
sCreate_Dense_Matrix(
&B, n, 1, b, n,
SLU_DN, /* Fortran-type column-wise storage */
SLU_S, /* floats */
SLU_GE /* general */
);
/* Step 3: get permutation matrix
* ------------------------------
* com_perm: 0 -> no re-ordering
* 1 -> re-ordering for A^t.A
* 2 -> re-ordering for A^t+A
* 3 -> approximate minimum degree ordering
*/
get_perm_c(do_perm ? 3 : 0, &A, perm);
/* Step 4: call SuperLU main routine
* ---------------------------------
*/
set_default_options(&options);
options.ColPerm = MY_PERMC;
StatInit(&stat);
sgssv(&options, &A, perm, perm_r, &L, &U, &B, &stat, &info);
/* Step 5: get the solution
* ------------------------
* Fortran-type column-wise storage
*/
vals = (DNformat*)B.Store;
rvals = (float*)(vals->nzval);
if(info == 0) {
for(i = 0; i < n; i++){
x[i] = rvals[i];
}
}
/* Step 6: cleanup
* ---------------
*/
/*
* For these two ones, only the "store" structure
* needs to be deallocated (the arrays have been allocated
* by us).
*/
Destroy_SuperMatrix_Store(&A);
Destroy_SuperMatrix_Store(&B);
/*
* These ones need to be fully deallocated (they have been
* allocated by SuperLU).
*/
Destroy_SuperNode_Matrix(&L);
Destroy_CompCol_Matrix(&U);
StatFree(&stat);
__NL_DELETE_ARRAY(xa);
__NL_DELETE_ARRAY(rhs);
__NL_DELETE_ARRAY(a);
__NL_DELETE_ARRAY(asub);
__NL_DELETE_ARRAY(perm_r);
__NL_DELETE_ARRAY(perm);
return (info == 0);
}
int main ( int argc, char *argv[] )
/**********************************************************************/
/*
Purpose:
SUPER_LU_S2 solves a symmetric sparse system read from a file.
Discussion:
The sparse matrix is stored in a file using the Harwell-Boeing
sparse matrix format. The file should be assigned to the standard
input of this program. For instance, if the matrix is stored
in the file "g10_rua.txt", the execution command might be:
super_lu_s2 < g10_rua.txt
Modified:
25 April 2004
Reference:
James Demmel, John Gilbert, Xiaoye Li,
SuperLU Users's Guide,
Sections 1 and 2.
Local parameters:
SuperMatrix L, the computed L factor.
int *perm_c, the column permutation vector.
int *perm_r, the row permutations from partial pivoting.
SuperMatrix U, the computed U factor.
*/
{
SuperMatrix A;
NCformat *Astore;
float *a;
int *asub;
SuperMatrix B;
int info;
SuperMatrix L;
int ldx;
SCformat *Lstore;
int m;
mem_usage_t mem_usage;
int n;
int nnz;
int nrhs;
superlu_options_t options;
int *perm_c;
int *perm_r;
float *rhs;
float *sol;
SuperLUStat_t stat;
SuperMatrix U;
NCformat *Ustore;
int *xa;
float *xact;
/*
Say hello.
*/
printf ( "\n" );
printf ( "SUPER_LU_S2:\n" );
printf ( " Read a symmetric sparse matrix A from standard input,\n");
printf ( " stored in Harwell-Boeing Sparse Matrix format.\n" );
printf ( "\n" );
printf ( " Solve a linear system A * X = B.\n" );
/*
Set the default input options:
options.Fact = DOFACT;
options.Equil = YES;
options.ColPerm = COLAMD;
options.DiagPivotThresh = 1.0;
options.Trans = NOTRANS;
options.IterRefine = NOREFINE;
options.SymmetricMode = NO;
options.PivotGrowth = NO;
options.ConditionNumber = NO;
options.PrintStat = YES;
*/
set_default_options ( &options );
/*
Now we modify the default options to use the symmetric mode.
*/
options.SymmetricMode = YES;
options.ColPerm = MMD_AT_PLUS_A;
options.DiagPivotThresh = 0.001;
/*
Read the matrix in Harwell-Boeing format.
*/
sreadhb ( &m, &n, &nnz, &a, &asub, &xa );
/*
Create storage for a compressed column matrix.
*/
sCreate_CompCol_Matrix ( &A, m, n, nnz, a, asub, xa, SLU_NC, SLU_S, SLU_GE );
Astore = A.Store;
printf ( "\n" );
printf ( " Dimension %dx%d; # nonzeros %d\n", A.nrow, A.ncol, Astore->nnz );
/*
Set up the right hand side.
*/
nrhs = 1;
rhs = floatMalloc ( m * nrhs );
if ( !rhs )
{
ABORT ( " Malloc fails for rhs[]." );
}
sCreate_Dense_Matrix ( &B, m, nrhs, rhs, m, SLU_DN, SLU_S, SLU_GE );
xact = floatMalloc ( n * nrhs );
if ( !xact )
{
ABORT ( " Malloc fails for rhs[]." );
}
ldx = n;
sGenXtrue ( n, nrhs, xact, ldx );
sFillRHS ( options.Trans, nrhs, xact, ldx, &A, &B );
perm_c = intMalloc ( n );
if ( !perm_c )
{
ABORT ( "Malloc fails for perm_c[]." );
}
perm_r = intMalloc ( m );
if ( !perm_r )
{
ABORT ( "Malloc fails for perm_r[]." );
}
/*
Initialize the statistics variables.
*/
StatInit ( &stat );
/*
Call SGSSV to factor the matrix and solve the linear system.
*/
sgssv ( &options, &A, perm_c, perm_r, &L, &U, &B, &stat, &info );
if ( info == 0 )
{
/*
To conveniently access the solution matrix, you need to get a pointer to it.
*/
sol = (float*) ((DNformat*) B.Store)->nzval;
/*
Compute the infinity norm of the error.
*/
sinf_norm_error ( nrhs, &B, xact );
Lstore = (SCformat *) L.Store;
Ustore = (NCformat *) U.Store;
printf ( "\n" );
printf ( " Number of nonzeros in factor L = %d\n", Lstore->nnz );
printf ( " Number of nonzeros in factor U = %d\n", Ustore->nnz );
printf ( " Number of nonzeros in L+U = %d\n",
Lstore->nnz + Ustore->nnz - n );
sQuerySpace ( &L, &U, &mem_usage );
printf ( "\n" );
printf ( " L\\U MB %.3f\ttotal MB needed %.3f\texpansions %d\n",
mem_usage.for_lu/1e6, mem_usage.total_needed/1e6,
mem_usage.expansions);
}
else
{
printf ( "\n" );
printf ( " SGSSV error returns INFO= %d\n", info );
if ( info <= n )
{
sQuerySpace ( &L, &U, &mem_usage );
printf ( "\n" );
printf (" L\\U MB %.3f\ttotal MB needed %.3f\texpansions %d\n",
mem_usage.for_lu/1e6, mem_usage.total_needed/1e6,
mem_usage.expansions );
}
}
if ( options.PrintStat )
{
StatPrint ( &stat );
}
StatFree ( &stat );
/*
Free the memory.
*/
SUPERLU_FREE ( rhs );
SUPERLU_FREE ( xact );
SUPERLU_FREE ( perm_r );
SUPERLU_FREE ( perm_c );
Destroy_CompCol_Matrix ( &A );
Destroy_SuperMatrix_Store ( &B );
Destroy_SuperNode_Matrix ( &L );
Destroy_CompCol_Matrix ( &U );
/*
Say goodbye.
*/
printf ( "\n" );
printf ( "SUPER_LU_S2:\n" );
printf ( " Normal end of execution.\n");
return 0;
}